Research map / Architectures and efficiency / CNN design
ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Revisits the texture-bias claim for ImageNet CNNs by replacing Geirhos et al.'s cue-conflict test with a suppression-based framework (patch shuffle/rotation for shape, bilateral/Gaussian for texture, grayscale/channel shuffle for color) and compares humans, CNNs, and ViTs across CV, medical imaging, and remote sensing.
Problem
Whether ImageNet-trained CNNs are inherently texture-biased, as claimed by the cue-conflict experiment of Geirhos et al., which conflates salience with reliance and has design confounds.
Approach
Measure feature reliance by accuracy drop under targeted suppression of shape, texture, or color rather than forced-choice preference on style-transfer cue-conflict images; apply the same framework to humans, many architectures, and three domains.
Key results
- Under local shape suppression, ResNet50-standard retains only 28% of original performance versus 76% for humans, while under texture suppression it retains 80%.
- Modern training raises local-shape retention for ResNet50-sota to 62% and ConvNeXtV2 to 65%, and CLIP ViT most closely matches humans (0.758 local shape).
- Domain analysis: CV models rely most on local shape, MI models depend more on color, and RS models show the highest texture reliance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use modern training strategy instead of standard training | ResNet50 · ImageNet16-like, local shape suppression | relative accuracy: 0.276 -> 0.618 | Helped |
| use ConvNeXtV2 instead of ResNet50-standard | ImageNet16-like model · ImageNet16-like, local shape suppression | relative accuracy: 0.276 -> 0.647 | Helped |
| use vision-language-supervised CLIP ViT rather than standard CNNs | ImageNet16-like model · ImageNet16-like, local shape suppression | relative accuracy: 0.276 -> 0.758 | Helped |
| suppress texture rather than local shape | ResNet50-standard · ImageNet16-like | relative accuracy: 0.276 -> 0.795 | Helped |
| use argmax instead of summed softmax with a 0.5 threshold (ablation) | ResNet50-standard · ImageNet16-like, local shape suppression | relative accuracy: 0.276 -> 0.361 | Helped |
| use argmax instead of summed softmax with a 0.5 threshold (ablation) | ResNet50-sota · ImageNet16-like, local shape suppression | relative accuracy: 0.618 -> 0.665 | Helped |
| use argmax instead of summed softmax with a 0.5 threshold (ablation) | ConvNeXtV2 · ImageNet16-like, local shape suppression | relative accuracy: 0.647 -> 0.748 | Helped |
| use bilateral filtering rather than box or median blur for texture suppression | suppression evaluation transformations · ImageNet validation images | normalized shape score: 0.796 vs 0.363; 0.506 | Helped |
| increase texture suppression strength with Gaussian blur rather than bilateral filtering (ablation) | suppression evaluation transformations · ImageNet validation images | normalized texture score: 0.349 vs 0.521 | Helped |
| increase texture suppression strength with Gaussian blur rather than bilateral filtering (ablation) | suppression evaluation transformations · ImageNet validation images | normalized shape score: 0.796 -> 0.662 | Hurt |
| replace Patch Shuffle with Patch Rotation for shape suppression (ablation) | suppression evaluation transformations · ImageNet validation images | normalized shape score: 0.176 -> 0.293 | Hurt |
| fine-tune with ImageNet pretraining rather than train from scratch (ablation) | ResNet50 on computer-vision datasets · computer vision, feature suppression | relative accuracy under non-shape suppression | Helped |
Surprising findings
- ResNet50-standard is more sensitive to local shape suppression than to texture suppression, contradicting the texture-bias claim.
- Color suppression via grayscale causes notable drops for remote sensing models despite RGB-only imagery.
Limitations
- Suppression only reduces rather than eliminates features, and transformations may introduce artifacts (block edges, smoothing).
- Results with pretrained models may reflect similarity between suppression transformations and augmentations such as Cutout and Patch Shuffle.
- Human study uses a controlled forced-choice design with brief exposures and a limited set of 16 categories.
Experimental setup
Datasets: ImageNet-1K Caltech101 Flowers102 Oxford-IIIT-Pet STL10 MedMNIST-v2 UCMerced RSD46-WHU DeepGlobe PatternNet AID
Benchmarks: ImageNet16-like (human study) Caltech101 Flowers102 Oxford-IIIT-Pet STL10 MedMNIST-v2 UCMerced RSD46-WHU DeepGlobe PatternNet AID
Metrics: relative accuracy under suppression accuracy Local Variance High-Frequency Energy Edge-SSIM Gradient Correlation Cohen's d
Baselines: ResNet50-standard ResNet50-sota ConvNeXt ConvNeXtV2 ViT CLIP ViT Human observers
Builds on: Geirhos et al. cue-conflict experiment ImageNet-trained CNNs MedMNIST-v2
Closest papers by meaning
- The Origins and Prevalence of Texture Bias in Convolutional Neural Networks (NeurIPS 2020 · similarity 0.95)
- Shape or Texture: Understanding Discriminative Features in CNNs (ICLR 2021 · similarity 0.92)
- ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness (ICLR 2019 · similarity 0.92)
- BiasBed - Rigorous Texture Bias Evaluation (CVPR 2023 · similarity 0.90)
- Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models (NeurIPS 2025 · similarity 0.88)
- The Shape of Events: Edge-Based Inductive Biases via Cross-Domain Distillation (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.